Datasets:

dim

/

scitldr

like 0

We study the problem of semantic code repair, which can be broadly defined as automatically fixing non-syntactic bugs in source code. The majority of past work in semantic code repair assumed access to unit tests against which candidate repairs could be validated. In contrast, the goal here is to develop a strong statistical model to accurately predict both bug locations and exact fixes without access to information about the intended correct behavior of the program. Achieving such a goal requires a robust contextual repair model, which we train on a large corpus of real-world source code that has been augmented with synthetically injected bugs. Our framework adopts a two-stage approach where first a large set of repair candidates are generated by rule-based processors, and then these candidates are scored by a statistical model using a novel neural network architecture which we refer to as Share, Specialize, and Compete. Specifically, the architecture (1) generates a shared encoding of the source code using an RNN over the abstract syntax tree, (2) scores each candidate repair using specialized network modules, and (3) then normalizes these scores together so they can compete against one another in comparable probability space. We evaluate our model on a real-world test set gathered from GitHub containing four common categories of bugs. Our model is able to predict the exact correct repair 41% of the time with a single guess, compared to 13% accuracy for an attentional sequence-to-sequence model.

A neural architecture for scoring and ranking program repair candidates to perform semantic program repair statically without access to unit tests.

Deep networks were recently suggested to face the odds between accuracy (on clean natural images) and robustness (on adversarially perturbed images) (Tsipras et al., 2019). Such a dilemma is shown to be rooted in the inherently higher sample complexity (Schmidt et al., 2018) and/or model capacity (Nakkiran, 2019), for learning a high-accuracy and robust classifier. In view of that, give a classification task, growing the model capacity appears to help draw a win-win between accuracy and robustness, yet at the expense of model size and latency, therefore posing challenges for resource-constrained applications. Is it possible to co-design model accuracy, robustness and efficiency to achieve their triple wins? This paper studies multi-exit networks associated with input-adaptive efficient inference, showing their strong promise in achieving a “sweet point" in co-optimizing model accuracy, robustness, and efficiency. Our proposed solution, dubbed Robust Dynamic Inference Networks (RDI-Nets), allows for each input (either clean or adversarial) to adaptively choose one of the multiple output layers (early branches or the final one) to output its prediction. That multi-loss adaptivity adds new variations and flexibility to adversarial attacks and defenses, on which we present a systematical investigation. We show experimentally that by equipping existing backbones with such robust adaptive inference, the resulting RDI-Nets can achieve better accuracy and robustness, yet with over 30% computational savings, compared to the defended original models.

Is it possible to co-design model accuracy, robustness and efficiency to achieve their triple wins? Yes!

Although deep convolutional networks have achieved improved performance in many natural language tasks, they have been treated as black boxes because they are difficult to interpret. Especially, little is known about how they represent language in their intermediate layers. In an attempt to understand the representations of deep convolutional networks trained on language tasks, we show that individual units are selectively responsive to specific morphemes, words, and phrases, rather than responding to arbitrary and uninterpretable patterns. In order to quantitatively analyze such intriguing phenomenon, we propose a concept alignment method based on how units respond to replicated text. We conduct analyses with different architectures on multiple datasets for classification and translation tasks and provide new insights into how deep models understand natural language.

We show that individual units in CNN representations learned in NLP tasks are selectively responsive to specific natural language concepts.

We study the problem of building models that disentangle independent factors of variation. Such models encode features that can efficiently be used for classification and to transfer attributes between different images in image synthesis. As data we use a weakly labeled training set, where labels indicate what single factor has changed between two data samples, although the relative value of the change is unknown. This labeling is of particular interest as it may be readily available without annotation costs. We introduce an autoencoder model and train it through constraints on image pairs and triplets. We show the role of feature dimensionality and adversarial training theoretically and experimentally. We formally prove the existence of the reference ambiguity, which is inherently present in the disentangling task when weakly labeled data is used. The numerical value of a factor has different meaning in different reference frames. When the reference depends on other factors, transferring that factor becomes ambiguous. We demonstrate experimentally that the proposed model can successfully transfer attributes on several datasets, but show also cases when the reference ambiguity occurs.

It is a mostly theoretical paper that describes the challenges in disentangling factors of variation, using autoencoders and GAN.

In information retrieval, learning to rank constructs a machine-based ranking model which given a query, sorts the search results by their degree of relevance or importance to the query. Neural networks have been successfully applied to this problem, and in this paper, we propose an attention-based deep neural network which better incorporates different embeddings of the queries and search results with an attention-based mechanism. This model also applies a decoder mechanism to learn the ranks of the search results in a listwise fashion. The embeddings are trained with convolutional neural networks or the word2vec model. We demonstrate the performance of this model with image retrieval and text querying data sets.

learning to rank with several embeddings and attentions

Computational neuroscience aims to fit reliable models of in vivo neural activity and interpret them as abstract computations. Recent work has shown that functional diversity of neurons may be limited to that of relatively few cell types; other work has shown that incorporating constraints into artificial neural networks (ANNs) can improve their ability to mimic neural data. Here we develop an algorithm that takes as input recordings of neural activity and returns clusters of neurons by cell type and models of neural activity constrained by these clusters. The resulting models are both more predictive and more interpretable, revealing the contributions of functional cell types to neural computation and ultimately informing the design of future ANNs.

We developed an algorithm that takes as input recordings of neural activity and returns clusters of neurons by cell type and models of neural activity constrained by these clusters.

Graph Neural Networks (GNNs) are a powerful representational tool for solving problems on graph-structured inputs. In almost all cases so far, however, they have been applied to directly recovering a final solution from raw inputs, without explicit guidance on how to structure their problem-solving. Here, instead, we focus on learning in the space of algorithms: we train several state-of-the-art GNN architectures to imitate individual steps of classical graph algorithms, parallel (breadth-first search, Bellman-Ford) as well as sequential (Prim's algorithm). As graph algorithms usually rely on making discrete decisions within neighbourhoods, we hypothesise that maximisation-based message passing neural networks are best-suited for such objectives, and validate this claim empirically. We also demonstrate how learning in the space of algorithms can yield new opportunities for positive transfer between tasks---showing how learning a shortest-path algorithm can be substantially improved when simultaneously learning a reachability algorithm.

We supervise graph neural networks to imitate intermediate and step-wise outputs of classical graph algorithms, recovering highly favourable insights.

Prospection is an important part of how humans come up with new task plans, but has not been explored in depth in robotics. Predicting multiple task-level is a challenging problem that involves capturing both task semantics and continuous variability over the state of the world. Ideally, we would combine the ability of machine learning to leverage big data for learning the semantics of a task, while using techniques from task planning to reliably generalize to new environment. In this work, we propose a method for learning a model encoding just such a representation for task planning. We learn a neural net that encodes the k most likely outcomes from high level actions from a given world. Our approach creates comprehensible task plans that allow us to predict changes to the environment many time steps into the future. We demonstrate this approach via application to a stacking task in a cluttered environment, where the robot must select between different colored blocks while avoiding obstacles, in order to perform a task. We also show results on a simple navigation task. Our algorithm generates realistic image and pose predictions at multiple points in a given task.

We describe an architecture for generating diverse hypotheses for intermediate goals during robotic manipulation tasks.

Adaptive gradient algorithms perform gradient-based updates using the history of gradients and are ubiquitous in training deep neural networks. While adaptive gradient methods theory is well understood for minimization problems, the underlying factors driving their empirical success in min-max problems such as GANs remain unclear. In this paper, we aim at bridging this gap from both theoretical and empirical perspectives. First, we analyze a variant of Optimistic Stochastic Gradient (OSG) proposed in~\citep{daskalakis2017training} for solving a class of non-convex non-concave min-max problem and establish $O(\epsilon^{-4})$ complexity for finding $\epsilon$-first-order stationary point, in which the algorithm only requires invoking one stochastic first-order oracle while enjoying state-of-the-art iteration complexity achieved by stochastic extragradient method by~\citep{iusem2017extragradient}. Then we propose an adaptive variant of OSG named Optimistic Adagrad (OAdagrad) and reveal an \emph{improved} adaptive complexity $\widetilde{O}\left(\epsilon^{-\frac{2}{1-\alpha}}\right)$~\footnote{Here $\widetilde{O}(\cdot)$ compresses a logarithmic factor of $\epsilon$. }, where $\alpha$ characterizes the growth rate of the cumulative stochastic gradient and $0\leq \alpha\leq 1/2$. To the best of our knowledge, this is the first work for establishing adaptive complexity in non-convex non-concave min-max optimization. Empirically, our experiments show that indeed adaptive gradient algorithms outperform their non-adaptive counterparts in GAN training. Moreover, this observation can be explained by the slow growth rate of the cumulative stochastic gradient, as observed empirically.

This paper provides novel analysis of adaptive gradient algorithms for solving non-convex non-concave min-max problems as GANs, and explains the reason why adaptive gradient methods outperform its non-adaptive counterparts by empirical studies.

We consider the problem of unsupervised learning of a low dimensional, interpretable, latent state of a video containing a moving object. The problem of distilling dynamics from pixels has been extensively considered through the lens of graphical/state space models that exploit Markov structure for cheap computation and structured graphical model priors for enforcing interpretability on latent representations. We take a step towards extending these approaches by discarding the Markov structure; instead, repurposing the recently proposed Gaussian Process Prior Variational Autoencoder for learning sophisticated latent trajectories. We describe the model and perform experiments on a synthetic dataset and see that the model reliably reconstructs smooth dynamics exhibiting U-turns and loops. We also observe that this model may be trained without any beta-annealing or freeze-thaw of training parameters. Training is performed purely end-to-end on the unmodified evidence lower bound objective. This is in contrast to previous works, albeit for slightly different use cases, where application specific training tricks are often required.

We learn sohpisticated trajectories of an object purely from pixels with a toy video dataset by using a VAE structure with a Gaussian process prior.

Dreams and our ability to recall them are among the most puzzling questions in sleep research. Specifically, putative differences in brain network dynamics between individuals with high versus low dream recall rates, are still poorly understood. In this study, we addressed this question as a classification problem where we applied deep convolutional networks (CNN) to sleep EEG recordings to predict whether subjects belonged to the high or low dream recall group (HDR and LDR resp.). Our model achieves significant accuracy levels across all the sleep stages, thereby indicating subtle signatures of dream recall in the sleep microstructure. We also visualized the feature space to inspect the subject-specificity of the learned features, thus ensuring that the network captured population level differences. Beyond being the first study to apply deep learning to sleep EEG in order to classify HDR and LDR, guided backpropagation allowed us to visualize the most discriminant features in each sleep stage. The significance of these findings and future directions are discussed.

We investigate the neural basis of dream recall using convolutional neural network and feature visualization techniques, like tSNE and guided-backpropagation.

This paper considers multi-agent reinforcement learning (MARL) in networked system control. Specifically, each agent learns a decentralized control policy based on local observations and messages from connected neighbors. We formulate such a networked MARL (NMARL) problem as a spatiotemporal Markov decision process and introduce a spatial discount factor to stabilize the training of each local agent. Further, we propose a new differentiable communication protocol, called NeurComm, to reduce information loss and non-stationarity in NMARL. Based on experiments in realistic NMARL scenarios of adaptive traffic signal control and cooperative adaptive cruise control, an appropriate spatial discount factor effectively enhances the learning curves of non-communicative MARL algorithms, while NeurComm outperforms existing communication protocols in both learning efficiency and control performance.

This paper proposes a new formulation and a new communication protocol for networked multi-agent control problems

Variational Bayesian Inference is a popular methodology for approximating posterior distributions over Bayesian neural network weights. Recent work developing this class of methods has explored ever richer parameterizations of the approximate posterior in the hope of improving performance. In contrast, here we share a curious experimental finding that suggests instead restricting the variational distribution to a more compact parameterization. For a variety of deep Bayesian neural networks trained using Gaussian mean-field variational inference, we find that the posterior standard deviations consistently exhibit strong low-rank structure after convergence. This means that by decomposing these variational parameters into a low-rank factorization, we can make our variational approximation more compact without decreasing the models' performance. Furthermore, we find that such factorized parameterizations improve the signal-to-noise ratio of stochastic gradient estimates of the variational lower bound, resulting in faster convergence.

Mean field VB uses twice as many parameters; we tie variance parameters in mean field VB without any loss in ELBO, gaining speed and lower variance gradients.

Aspect extraction in online product reviews is a key task in sentiment analysis and opinion mining. Training supervised neural networks for aspect extraction is not possible when ground truth aspect labels are not available, while the unsupervised neural topic models fail to capture the particular aspects of interest. In this work, we propose a weakly supervised approach for training neural networks for aspect extraction in cases where only a small set of seed words, i.e., keywords that describe an aspect, are available. Our main contributions are as follows. First, we show that current weakly supervised networks fail to leverage the predictive power of the available seed words by comparing them to a simple bag-of-words classifier. Second, we propose a distillation approach for aspect extraction where the seed words are considered by the bag-of-words classifier (teacher) and distilled to the parameters of a neural network (student). Third, we show that regularization encourages the student to consider non-seed words for classification and, as a result, the student outperforms the teacher, which only considers the seed words. Finally, we empirically show that our proposed distillation approach outperforms (by up to 34.4% in F1 score) previous weakly supervised approaches for aspect extraction in six domains of Amazon product reviews.

We effectively leverage a few keywords as weak supervision for training neural networks for aspect extraction.

Forming perceptual groups and individuating objects in visual scenes is an essential step towards visual intelligence. This ability is thought to arise in the brain from computations implemented by bottom-up, horizontal, and top-down connections between neurons. However, the relative contributions of these connections to perceptual grouping are poorly understood. We address this question by systematically evaluating neural network architectures featuring combinations of these connections on two synthetic visual tasks, which stress low-level "Gestalt" vs. high-level object cues for perceptual grouping. We show that increasing the difficulty of either task strains learning for networks that rely solely on bottom-up processing. Horizontal connections resolve this limitation on tasks with Gestalt cues by supporting incremental spatial propagation of activities, whereas top-down connections rescue learning on tasks with high-level object cues by modifying coarse predictions about the position of the target object. Our findings dissociate the computational roles of bottom-up, horizontal and top-down connectivity, and demonstrate how a model featuring all of these interactions can more flexibly learn to form perceptual groups.

Horizontal and top-down feedback connections are responsible for complementary perceptual grouping strategies in biological and recurrent vision systems.

Generative adversarial networks have seen rapid development in recent years and have led to remarkable improvements in generative modelling of images. However, their application in the audio domain has received limited attention, and autoregressive models, such as WaveNet, remain the state of the art in generative modelling of audio signals such as human speech. To address this paucity, we introduce GAN-TTS, a Generative Adversarial Network for Text-to-Speech. Our architecture is composed of a conditional feed-forward generator producing raw speech audio, and an ensemble of discriminators which operate on random windows of different sizes. The discriminators analyse the audio both in terms of general realism, as well as how well the audio corresponds to the utterance that should be pronounced. To measure the performance of GAN-TTS, we employ both subjective human evaluation (MOS - Mean Opinion Score), as well as novel quantitative metrics (Fréchet DeepSpeech Distance and Kernel DeepSpeech Distance), which we find to be well correlated with MOS. We show that GAN-TTS is capable of generating high-fidelity speech with naturalness comparable to the state-of-the-art models, and unlike autoregressive models, it is highly parallelisable thanks to an efficient feed-forward generator. Listen to GAN-TTS reading this abstract at http://tiny.cc/gantts.

We introduce GAN-TTS, a Generative Adversarial Network for Text-to-Speech, which achieves Mean Opinion Score (MOS) 4.2.

This paper proposes a Pruning in Training (PiT) framework of learning to reduce the parameter size of networks. Different from existing works, our PiT framework employs the sparse penalties to train networks and thus help rank the importance of weights and filters. Our PiT algorithms can directly prune the network without any fine-tuning. The pruned networks can still achieve comparable performance to the original networks. In particular, we introduce the (Group) Lasso-type Penalty (L-P /GL-P), and (Group) Split LBI Penalty (S-P / GS-P) to regularize the networks, and a pruning strategy proposed is used in help prune the network. We conduct the extensive experiments on MNIST, Cifar-10, and miniImageNet. The results validate the efficacy of our proposed methods. Remarkably, on MNIST dataset, our PiT framework can save 17.5% parameter size of LeNet-5, which achieves the 98.47% recognition accuracy.

we propose an algorithm of learning to prune network by enforcing structure sparsity penalties

We first pose the Unsupervised Continual Learning (UCL) problem: learning salient representations from a non-stationary stream of unlabeled data in which the number of object classes varies with time. Given limited labeled data just before inference, those representations can also be associated with specific object types to perform classification. To solve the UCL problem, we propose an architecture that involves a single module, called Self-Taught Associative Memory (STAM), which loosely models the function of a cortical column in the mammalian brain. Hierarchies of STAM modules learn based on a combination of Hebbian learning, online clustering, detection of novel patterns and forgetting outliers, and top-down predictions. We illustrate the operation of STAMs in the context of learning handwritten digits in a continual manner with only 3-12 labeled examples per class. STAMs suggest a promising direction to solve the UCL problem without catastrophic forgetting.

We introduce unsupervised continual learning (UCL) and a neuro-inspired architecture that solves the UCL problem.

Recent advances have made it possible to create deep complex-valued neural networks. Despite this progress, the potential power of fully complex intermediate computations and representations has not yet been explored for many challenging learning problems. Building on recent advances, we propose a novel mechanism for extracting signals in the frequency domain. As a case study, we perform audio source separation in the Fourier domain. Our extraction mechanism could be regarded as a local ensembling method that combines a complex-valued convolutional version of Feature-Wise Linear Modulation (FiLM) and a signal averaging operation. We also introduce a new explicit amplitude and phase-aware loss, which is scale and time invariant, taking into account the complex-valued components of the spectrogram. Using the Wall Street Journal Dataset, we compare our phase-aware loss to several others that operate both in the time and frequency domains and demonstrate the effectiveness of our proposed signal extraction method and proposed loss. When operating in the complex-valued frequency domain, our deep complex-valued network substantially outperforms its real-valued counterparts even with half the depth and a third of the parameters. Our proposed mechanism improves significantly deep complex-valued networks' performance and we demonstrate the usefulness of its regularizing effect.

New Signal Extraction Method in the Fourier Domain

It is challenging to disentangle an object into two orthogonal spaces of content and style since each can influence the visual observation in a different and unpredictable way. It is rare for one to have access to a large number of data to help separate the influences. In this paper, we present a novel framework to learn this disentangled representation in a completely unsupervised manner. We address this problem in a two-branch Autoencoder framework. For the structural content branch, we project the latent factor into a soft structured point tensor and constrain it with losses derived from prior knowledge. This encourages the branch to distill geometry information. Another branch learns the complementary style information. The two branches form an effective framework that can disentangle object's content-style representation without any human annotation. We evaluate our approach on four image datasets, on which we demonstrate the superior disentanglement and visual analogy quality both in synthesized and real-world data. We are able to generate photo-realistic images with 256x256 resolution that are clearly disentangled in content and style.

We present a novel framework to learn the disentangled representation of content and style in a completely unsupervised manner.

We develop the Y-learner for estimating heterogeneous treatment effects in experimental and observational studies. The Y-learner is designed to leverage the abilities of neural networks to optimize multiple objectives and continually update, which allows for better pooling of underlying feature information between treatment and control groups. We evaluate the Y-learner on three test problems: (1) A set of six simulated data benchmarks from the literature. (2) A real-world large-scale experiment on voter persuasion. (3) A task from the literature that estimates artificially generated treatment effects on MNIST didgits. The Y-learner achieves state of the art results on two of the three tasks. On the MNIST task, it gets the second best results.

We develop a CATE estimation strategy that takes advantage some of the intriguing properties of neural networks.

With the rapid proliferation of IoT devices, our cyberspace is nowadays dominated by billions of low-cost computing nodes, which expose an unprecedented heterogeneity to our computing systems. Dynamic analysis, one of the most effective approaches to finding software bugs, has become paralyzed due to the lack of a generic emulator capable of running diverse previously-unseen firmware. In recent years, we have witnessed devastating security breaches targeting IoT devices. These security concerns have significantly hamstrung further evolution of IoT technology. In this work, we present Laelaps, a device emulator specifically designed to run diverse software on low-cost IoT devices. We do not encode into our emulator any specific information about a device. Instead, Laelaps infers the expected behavior of firmware via symbolic-execution-assisted peripheral emulation and generates proper inputs to steer concrete execution on the fly. This unique design feature makes Laelaps the first generic device emulator capable of running diverse firmware with no a priori knowledge about the target device. To demonstrate the capabilities of Laelaps, we deployed two popular dynamic analysis techniques---fuzzing testing and dynamic symbolic execution---on top of our emulator. We successfully identified both self-injected and real-world vulnerabilities.

Device-agnostic Firmware Execution

Deep neural models, such as convolutional and recurrent networks, achieve phenomenal results over spatial data such as images and text. However, when considering tabular data, gradient boosting of decision trees (GBDT) remains the method of choice. Aiming to bridge this gap, we propose \emph{deep neural forests} (DNF) -- a novel architecture that combines elements from decision trees as well as dense residual connections. We present the results of extensive empirical study in which we examine the performance of GBDTs, DNFs and (deep) fully-connected networks. These results indicate that DNFs achieve comparable results to GBDTs on tabular data, and open the door to end-to-end neural modeling of multi-modal data. To this end, we present a successful application of DNFs as part of a hybrid architecture for a multi-modal driving scene understanding classification task.

An architecture for tabular data, which emulates branches of decision trees and uses dense residual connectivity

Hyperparameter tuning is one of the most time-consuming workloads in deep learning. State-of-the-art optimizers, such as AdaGrad, RMSProp and Adam, reduce this labor by adaptively tuning an individual learning rate for each variable. Recently researchers have shown renewed interest in simpler methods like momentum SGD as they may yield better results. Motivated by this trend, we ask: can simple adaptive methods, based on SGD perform as well or better? We revisit the momentum SGD algorithm and show that hand-tuning a single learning rate and momentum makes it competitive with Adam. We then analyze its robustness to learning rate misspecification and objective curvature variation. Based on these insights, we design YellowFin, an automatic tuner for momentum and learning rate in SGD. YellowFin optionally uses a negative-feedback loop to compensate for the momentum dynamics in asynchronous settings on the fly. We empirically show YellowFin can converge in fewer iterations than Adam on ResNets and LSTMs for image recognition, language modeling and constituency parsing, with a speedup of up to $3.28$x in synchronous and up to $2.69$x in asynchronous settings.

YellowFin is an SGD based optimizer with both momentum and learning rate adaptivity.

Robustness and security of machine learning (ML) systems are intertwined, wherein a non-robust ML system (classifiers, regressors, etc.) can be subject to attacks using a wide variety of exploits. With the advent of scalable deep learning methodologies, a lot of emphasis has been put on the robustness of supervised, unsupervised and reinforcement learning algorithms. Here, we study the robustness of the latent space of a deep variational autoencoder (dVAE), an unsupervised generative framework, to show that it is indeed possible to perturb the latent space, flip the class predictions and keep the classification probability approximately equal before and after an attack. This means that an agent that looks at the outputs of a decoder would remain oblivious to an attack.

Adversarial attacks on the latent space of variational autoencoders to change the semantic meaning of inputs

Graph-based dependency parsing consists of two steps: first, an encoder produces a feature representation for each parsing substructure of the input sentence, which is then used to compute a score for the substructure; and second, a decoder} finds the parse tree whose substructures have the largest total score. Over the past few years, powerful neural techniques have been introduced into the encoding step which substantially increases parsing accuracies. However, advanced decoding techniques, in particular high-order decoding, have seen a decline in usage. It is widely believed that contextualized features produced by neural encoders can help capture high-order decoding information and hence diminish the need for a high-order decoder. In this paper, we empirically evaluate the combinations of different neural and non-neural encoders with first- and second-order decoders and provide a comprehensive analysis about the effectiveness of these combinations with varied training data sizes. We find that: first, when there is large training data, a strong neural encoder with first-order decoding is sufficient to achieve high parsing accuracy and only slightly lags behind the combination of neural encoding and second-order decoding; second, with small training data, a non-neural encoder with a second-order decoder outperforms the other combinations in most cases.

An empirical study that examines the effectiveness of different encoder-decoder combinations for the task of dependency parsing

Meta-learning will be crucial to creating lifelong, generalizable AI. In practice, however, it is hard to define the meta-training task distribution that is used to train meta-learners. If made too small, tasks are too similar for a model to meaningfully generalize. If made too large, generalization becomes incredibly difficult. We argue that both problems can be alleviated by introducing a teacher model that controls the sequence of tasks that a meta-learner is trained on. This teacher model is incentivized to start the student meta-learner on simple tasks then adaptively increase task difficulty in response to student progress. While this approach has been previously studied in curriculum generation, our main contribution is in extending it to meta-learning.

Teacher that trains meta-learners like humans

Using higher order knowledge to reduce training data has become a popular research topic. However, the ability for available methods to draw effective decision boundaries is still limited: when training set is small, neural networks will be biased to certain labels. Based on this observation, we consider constraining output probability distribution as higher order domain knowledge. We design a novel algorithm that jointly optimizes output probability distribution on a clustered embedding space to make neural networks draw effective decision boundaries. While directly applying probability constraint is not effective, users need to provide additional very weak supervisions: mark some batches that have output distribution greatly differ from target probability distribution. We use experiments to empirically prove that our model can converge to an accuracy higher than other state-of-art semi-supervised learning models with less high quality labeled training examples.

We introduce an embedding space approach to constrain neural network output probability distribution.

We introduce a new type of deep contextualized word representation that models both (1) complex characteristics of word use (e.g., syntax and semantics), and (2) how these uses vary across linguistic contexts (i.e., to model polysemy). Our word vectors are learned functions of the internal states of a deep bidirectional language model (biLM), which is pretrained on a large text corpus. We show that these representations can be easily added to existing models and significantly improve the state of the art across six challenging NLP problems, including question answering, textual entailment and sentiment analysis. We also present an analysis showing that exposing the deep internals of the pretrained network is crucial, allowing downstream models to mix different types of semi-supervision signals.

We introduce a new type of deep contextualized word representation that significantly improves the state of the art for a range of challenging NLP tasks.

This work addresses the long-standing problem of robust event localization in the presence of temporally of misaligned labels in the training data. We propose a novel versatile loss function that generalizes a number of training regimes from standard fully-supervised cross-entropy to count-based weakly-supervised learning. Unlike classical models which are constrained to strictly fit the annotations during training, our soft localization learning approach relaxes the reliance on the exact position of labels instead. Training with this new loss function exhibits strong robustness to temporal misalignment of labels, thus alleviating the burden of precise annotation of temporal sequences. We demonstrate state-of-the-art performance against standard benchmarks in a number of challenging experiments and further show that robustness to label noise is not achieved at the expense of raw performance.

This work introduces a novel loss function for the robust training of temporal localization DNN in the presence of misaligned labels.

The driving force behind deep networks is their ability to compactly represent rich classes of functions. The primary notion for formally reasoning about this phenomenon is expressive efficiency, which refers to a situation where one network must grow unfeasibly large in order to replicate functions of another. To date, expressive efficiency analyses focused on the architectural feature of depth, showing that deep networks are representationally superior to shallow ones. In this paper we study the expressive efficiency brought forth by connectivity, motivated by the observation that modern networks interconnect their layers in elaborate ways. We focus on dilated convolutional networks, a family of deep models delivering state of the art performance in sequence processing tasks. By introducing and analyzing the concept of mixed tensor decompositions, we prove that interconnecting dilated convolutional networks can lead to expressive efficiency. In particular, we show that even a single connection between intermediate layers can already lead to an almost quadratic gap, which in large-scale settings typically makes the difference between a model that is practical and one that is not. Empirical evaluation demonstrates how the expressive efficiency of connectivity, similarly to that of depth, translates into gains in accuracy. This leads us to believe that expressive efficiency may serve a key role in developing new tools for deep network design.

We introduce the notion of mixed tensor decompositions, and use it to prove that interconnecting dilated convolutional networks boosts their expressive power.

We apply multi-task learning to image classification tasks on MNIST-like datasets. MNIST dataset has been referred to as the {\em drosophila} of machine learning and has been the testbed of many learning theories. The NotMNIST dataset and the FashionMNIST dataset have been created with the MNIST dataset as reference. In this work, we exploit these MNIST-like datasets for multi-task learning. The datasets are pooled together for learning the parameters of joint classification networks. Then the learned parameters are used as the initial parameters to retrain disjoint classification networks. The baseline recognition model are all-convolution neural networks. Without multi-task learning, the recognition accuracies for MNIST, NotMNIST and FashionMNIST are 99.56\%, 97.22\% and 94.32\% respectively. With multi-task learning to pre-train the networks, the recognition accuracies are respectively 99.70\%, 97.46\% and 95.25\%. The results re-affirm that multi-task learning framework, even with data with different genres, does lead to significant improvement.

multi-task learning works

Recent work has demonstrated that neural networks are vulnerable to adversarial examples, i.e., inputs that are almost indistinguishable from natural data and yet classified incorrectly by the network. To address this problem, we study the adversarial robustness of neural networks through the lens of robust optimization. This approach provides us with a broad and unifying view on much prior work on this topic. Its principled nature also enables us to identify methods for both training and attacking neural networks that are reliable and, in a certain sense, universal. In particular, they specify a concrete security guarantee that would protect against a well-defined class of adversaries. These methods let us train networks with significantly improved resistance to a wide range of adversarial attacks. They also suggest robustness against a first-order adversary as a natural security guarantee. We believe that robustness against such well-defined classes of adversaries is an important stepping stone towards fully resistant deep learning models.

We provide a principled, optimization-based re-look at the notion of adversarial examples, and develop methods that produce models that are adversarially robust against a wide range of adversaries.

In recent years there has been a rapid increase in classification methods on graph structured data. Both in graph kernels and graph neural networks, one of the implicit assumptions of successful state-of-the-art models was that incorporating graph isomorphism features into the architecture leads to better empirical performance. However, as we discover in this work, commonly used data sets for graph classification have repeating instances which cause the problem of isomorphism bias, i.e. artificially increasing the accuracy of the models by memorizing target information from the training set. This prevents fair competition of the algorithms and raises a question of the validity of the obtained results. We analyze 54 data sets, previously extensively used for graph-related tasks, on the existence of isomorphism bias, give a set of recommendations to machine learning practitioners to properly set up their models, and open source new data sets for the future experiments.

Many graph classification data sets have duplicates, thus raising questions about generalization abilities and fair comparison of the models.

Imitation learning, followed by reinforcement learning algorithms, is a promising paradigm to solve complex control tasks sample-efficiently. However, learning from demonstrations often suffers from the covariate shift problem, which results in cascading errors of the learned policy. We introduce a notion of conservatively extrapolated value functions, which provably lead to policies with self-correction. We design an algorithm Value Iteration with Negative Sampling (VINS) that practically learns such value functions with conservative extrapolation. We show that VINS can correct mistakes of the behavioral cloning policy on simulated robotics benchmark tasks. We also propose the algorithm of using VINS to initialize a reinforcement learning algorithm, which is shown to outperform prior works in sample efficiency.

We introduce a notion of conservatively-extrapolated value functions, which provably lead to policies that can self-correct to stay close to the demonstration states, and learn them with a novel negative sampling technique.

A structured understanding of our world in terms of objects, relations, and hierarchies is an important component of human cognition. Learning such a structured world model from raw sensory data remains a challenge. As a step towards this goal, we introduce Contrastively-trained Structured World Models (C-SWMs). C-SWMs utilize a contrastive approach for representation learning in environments with compositional structure. We structure each state embedding as a set of object representations and their relations, modeled by a graph neural network. This allows objects to be discovered from raw pixel observations without direct supervision as part of the learning process. We evaluate C-SWMs on compositional environments involving multiple interacting objects that can be manipulated independently by an agent, simple Atari games, and a multi-object physics simulation. Our experiments demonstrate that C-SWMs can overcome limitations of models based on pixel reconstruction and outperform typical representatives of this model class in highly structured environments, while learning interpretable object-based representations.

Contrastively-trained Structured World Models (C-SWMs) learn object-oriented state representations and a relational model of an environment from raw pixel input.

Neural machine translation (NMT) models learn representations containing substantial linguistic information. However, it is not clear if such information is fully distributed or if some of it can be attributed to individual neurons. We develop unsupervised methods for discovering important neurons in NMT models. Our methods rely on the intuition that different models learn similar properties, and do not require any costly external supervision. We show experimentally that translation quality depends on the discovered neurons, and find that many of them capture common linguistic phenomena. Finally, we show how to control NMT translations in predictable ways, by modifying activations of individual neurons.

Unsupervised methods for finding, analyzing, and controlling important neurons in NMT

Computations for the softmax function in neural network models are expensive when the number of output classes is large. This can become a significant issue in both training and inference for such models. In this paper, we present Doubly Sparse Softmax (DS-Softmax), Sparse Mixture of Sparse of Sparse Experts, to improve the efficiency for softmax inference. During training, our method learns a two-level class hierarchy by dividing entire output class space into several partially overlapping experts. Each expert is responsible for a learned subset of the output class space and each output class only belongs to a small number of those experts. During inference, our method quickly locates the most probable expert to compute small-scale softmax. Our method is learning-based and requires no knowledge of the output class partition space a priori. We empirically evaluate our method on several real-world tasks and demonstrate that we can achieve significant computation reductions without loss of

We present doubly sparse softmax, the sparse mixture of sparse of sparse experts, to improve the efficiency for softmax inference through exploiting the two-level overlapping hierarchy.

Our work presents empirical evidence that layer rotation, i.e. the evolution across training of the cosine distance between each layer's weight vector and its initialization, constitutes an impressively consistent indicator of generalization performance. Compared to previously studied indicators of generalization, we show that layer rotation has the additional benefit of being easily monitored and controlled, as well as having a network-independent optimum: the training procedures during which all layers' weights reach a cosine distance of 1 from their initialization consistently outperform other configurations -by up to 20% test accuracy. Finally, our results also suggest that the study of layer rotation can provide a unified framework to explain the impact of weight decay and adaptive gradient methods on generalization.

This paper presents empirical evidence supporting the discovery of an indicator of generalization: the evolution across training of the cosine distance between each layer's weight vector and its initialization.

Models of code can learn distributed representations of a program's syntax and semantics to predict many non-trivial properties of a program. Recent state-of-the-art models leverage highly structured representations of programs, such as trees, graphs and paths therein (e.g. data-flow relations), which are precise and abundantly available for code. This provides a strong inductive bias towards semantically meaningful relations, yielding more generalizable representations than classical sequence-based models. Unfortunately, these models primarily rely on graph-based message passing to represent relations in code, which makes them de facto local due to the high cost of message-passing steps, quite in contrast to modern, global sequence-based models, such as the Transformer. In this work, we bridge this divide between global and structured models by introducing two new hybrid model families that are both global and incorporate structural bias: Graph Sandwiches, which wrap traditional (gated) graph message-passing layers in sequential message-passing layers; and Graph Relational Embedding Attention Transformers (GREAT for short), which bias traditional Transformers with relational information from graph edge types. By studying a popular, non-trivial program repair task, variable-misuse identification, we explore the relative merits of traditional and hybrid model families for code representation. Starting with a graph-based model that already improves upon the prior state-of-the-art for this task by 20%, we show that our proposed hybrid models improve an additional 10-15%, while training both faster and using fewer parameters.

Models of source code that combine global and structural features learn more powerful representations of programs.

Recurrent neural networks (RNNs) are particularly well-suited for modeling long-term dependencies in sequential data, but are notoriously hard to train because the error backpropagated in time either vanishes or explodes at an exponential rate. While a number of works attempt to mitigate this effect through gated recurrent units, skip-connections, parametric constraints and design choices, we propose a novel incremental RNN (iRNN), where hidden state vectors keep track of incremental changes, and as such approximate state-vector increments of Rosenblatt's (1962) continuous-time RNNs. iRNN exhibits identity gradients and is able to account for long-term dependencies (LTD). We show that our method is computationally efficient overcoming overheads of many existing methods that attempt to improve RNN training, while suffering no performance degradation. We demonstrate the utility of our approach with extensive experiments and show competitive performance against standard LSTMs on LTD and other non-LTD tasks.

Incremental-RNNs resolves exploding/vanishing gradient problem by updating state vectors based on difference between previous state and that predicted by an ODE.

Recent empirical results on over-parameterized deep networks are marked by a striking absence of the classic U-shaped test error curve: test error keeps decreasing in wider networks. Researchers are actively working on bridging this discrepancy by proposing better complexity measures. Instead, we directly measure prediction bias and variance for four classification and regression tasks on modern deep networks. We find that both bias and variance can decrease as the number of parameters grows. Qualitatively, the phenomenon persists over a number of gradient-based optimizers. To better understand the role of optimization, we decompose the total variance into variance due to training set sampling and variance due to initialization. Variance due to initialization is significant in the under-parameterized regime. In the over-parameterized regime, total variance is much lower and dominated by variance due to sampling. We provide theoretical analysis in a simplified setting that is consistent with our empirical findings.

We provide evidence against classical claims about the bias-variance tradeoff and propose a novel decomposition for variance.

Real world images often contain large amounts of private / sensitive information that should be carefully protected without reducing their utilities. In this paper, we propose a privacy-preserving deep learning framework with a learnable ob- fuscator for the image classification task. Our framework consists of three mod- els: learnable obfuscator, classifier and reconstructor. The learnable obfuscator is used to remove the sensitive information in the images and extract the feature maps from them. The reconstructor plays the role as an attacker, which tries to recover the image from the feature maps extracted by the obfuscator. In order to best protect users’ privacy in images, we design an adversarial training methodol- ogy for our framework to optimize the obfuscator. Through extensive evaluations on real world datasets, both the numerical metrics and the visualization results demonstrate that our framework is qualified to protect users’ privacy and achieve a relatively high accuracy on the image classification task.

We proposed a novel deep learning image classification framework that can both accurately classify images and protect users' privacy.

Bitcoin is a virtual coinage system that enables users to trade virtually free of a central trusted authority. All transactions on the Bitcoin blockchain are publicly available for viewing, yet as Bitcoin is built mainly for security it’s original structure does not allow for direct analysis of address transactions. Existing analysis methods of the Bitcoin blockchain can be complicated, computationally expensive or inaccurate. We propose a computationally efficient model to analyze bitcoin blockchain addresses and allow for their use with existing machine learning algorithms. We compare our approach against Multi Level Sequence Learners (MLSLs), one of the best performing models on bitcoin address data.

a 2vec model for cryptocurrency transaction graphs

Despite remarkable empirical success, the training dynamics of generative adversarial networks (GAN), which involves solving a minimax game using stochastic gradients, is still poorly understood. In this work, we analyze last-iterate convergence of simultaneous gradient descent (simGD) and its variants under the assumption of convex-concavity, guided by a continuous-time analysis with differential equations. First, we show that simGD, as is, converges with stochastic sub-gradients under strict convexity in the primal variable. Second, we generalize optimistic simGD to accommodate an optimism rate separate from the learning rate and show its convergence with full gradients. Finally, we present anchored simGD, a new method, and show convergence with stochastic subgradients.

Convergence proof of stochastic sub-gradients method and variations on convex-concave minimax problems

Small spacecraft now have precise attitude control systems available commercially, allowing them to slew in 3 degrees of freedom, and capture images within short notice. When combined with appropriate software, this agility can significantly increase response rate, revisit time and coverage. In prior work, we have demonstrated an algorithmic framework that combines orbital mechanics, attitude control and scheduling optimization to plan the time-varying, full-body orientation of agile, small spacecraft in a constellation. The proposed schedule optimization would run at the ground station autonomously, and the resultant schedules uplinked to the spacecraft for execution. The algorithm is generalizable over small steerable spacecraft, control capability, sensor specs, imaging requirements, and regions of interest. In this article, we modify the algorithm to run onboard small spacecraft, such that the constellation can make time-sensitive decisions to slew and capture images autonomously, without ground control. We have developed a communication module based on Delay/Disruption Tolerant Networking (DTN) for onboard data management and routing among the satellites, which will work in conjunction with the other modules to optimize the schedule of agile communication and steering. We then apply this preliminary framework on representative constellations to simulate targeted measurements of episodic precipitation events and subsequent urban floods. The command and control efficiency of our agile algorithm is compared to non-agile (11.3x improvement) and non-DTN (21% improvement) constellations.

We propose an algorithmic framework to schedule constellations of small spacecraft with 3-DOF re-orientation capabilities, networked with inter-sat links.

Importance sampling (IS) is a standard Monte Carlo (MC) tool to compute information about random variables such as moments or quantiles with unknown distributions. IS is asymptotically consistent as the number of MC samples, and hence deltas (particles) that parameterize the density estimate, go to infinity. However, retaining infinitely many particles is intractable. We propose a scheme for only keeping a \emph{finite representative subset} of particles and their augmented importance weights that is \emph{nearly consistent}. To do so in {an online manner}, we approximate importance sampling in two ways. First, we replace the deltas by kernels, yielding kernel density estimates (KDEs). Second, we sequentially project KDEs onto nearby lower-dimensional subspaces. We characterize the asymptotic bias of this scheme as determined by a compression parameter and kernel bandwidth, which yields a tunable tradeoff between consistency and memory. In experiments, we observe a favorable tradeoff between memory and accuracy, providing for the first time near-consistent compressions of arbitrary posterior distributions.

We proposed a novel compressed kernelized importance sampling algorithm.

We study the following three fundamental problems about ridge regression: (1) what is the structure of the estimator? (2) how to correctly use cross-validation to choose the regularization parameter? and (3) how to accelerate computation without losing too much accuracy? We consider the three problems in a unified large-data linear model. We give a precise representation of ridge regression as a covariance matrix-dependent linear combination of the true parameter and the noise. We study the bias of $K$-fold cross-validation for choosing the regularization parameter, and propose a simple bias-correction. We analyze the accuracy of primal and dual sketching for ridge regression, showing they are surprisingly accurate. Our results are illustrated by simulations and by analyzing empirical data.

We study the structure of ridge regression in a high-dimensional asymptotic framework, and get insights about cross-validation and sketching.

Attention mechanisms have advanced the state of the art in several machine learning tasks. Despite significant empirical gains, there is a lack of theoretical analyses on understanding their effectiveness. In this paper, we address this problem by studying the landscape of population and empirical loss functions of attention-based neural networks. Our results show that, under mild assumptions, every local minimum of a two-layer global attention model has low prediction error, and attention models require lower sample complexity than models not employing attention. We then extend our analyses to the popular self-attention model, proving that they deliver consistent predictions with a more expressive class of functions. Additionally, our theoretical results provide several guidelines for designing attention mechanisms. Our findings are validated with satisfactory experimental results on MNIST and IMDB reviews dataset.

We analyze the loss landscape of neural networks with attention and explain why attention is helpful in training neural networks to achieve good performance.

Recent advances in deep learning techniques has shown the usefulness of the deep neural networks in extracting features required to perform the task at hand. However, these features learnt are in particular helpful only for the initial task. This is due to the fact that the features learnt are very task specific and does not capture the most general and task agnostic features of the input. In fact the way humans are seen to learn is by disentangling features which task agnostic. This indicates that leaning task agnostic features by disentangling only the most informative features from the input data. Recently Variational Auto-Encoders (VAEs) have shown to be the de-facto models to capture the latent variables in a generative sense. As these latent features can be represented as continuous and/or discrete variables, this indicates us to use VAE with a mixture of continuous and discrete variables for the latent space. We achieve this by performing our experiments using a modified version of joint-vae to learn the disentangled features.

Mixture Model for Neural Disentanglement

To improve how neural networks function it is crucial to understand their learning process. The information bottleneck theory of deep learning proposes that neural networks achieve good generalization by compressing their representations to disregard information that is not relevant to the task. However, empirical evidence for this theory is conflicting, as compression was only observed when networks used saturating activation functions. In contrast, networks with non-saturating activation functions achieved comparable levels of task performance but did not show compression. In this paper we developed more robust mutual information estimation techniques, that adapt to hidden activity of neural networks and produce more sensitive measurements of activations from all functions, especially unbounded functions. Using these adaptive estimation techniques, we explored compression in networks with a range of different activation functions. With two improved methods of estimation, firstly, we show that saturation of the activation function is not required for compression, and the amount of compression varies between different activation functions. We also find that there is a large amount of variation in compression between different network initializations. Secondary, we see that L2 regularization leads to significantly increased compression, while preventing overfitting. Finally, we show that only compression of the last layer is positively correlated with generalization.

We developed robust mutual information estimates for DNNs and used them to observe compression in networks with non-saturating activation functions

In this work, we address the problem of musical timbre transfer, where the goal is to manipulate the timbre of a sound sample from one instrument to match another instrument while preserving other musical content, such as pitch, rhythm, and loudness. In principle, one could apply image-based style transfer techniques to a time-frequency representation of an audio signal, but this depends on having a representation that allows independent manipulation of timbre as well as high-quality waveform generation. We introduce TimbreTron, a method for musical timbre transfer which applies “image” domain style transfer to a time-frequency representation of the audio signal, and then produces a high-quality waveform using a conditional WaveNet synthesizer. We show that the Constant Q Transform (CQT) representation is particularly well-suited to convolutional architectures due to its approximate pitch equivariance. Based on human perceptual evaluations, we confirmed that TimbreTron recognizably transferred the timbre while otherwise preserving the musical content, for both monophonic and polyphonic samples. We made an accompanying demo video here: https://www.cs.toronto.edu/~huang/TimbreTron/index.html which we strongly encourage you to watch before reading the paper.

We present the TimbreTron, a pipeline for perfoming high-quality timbre transfer on musical waveforms using CQT-domain style transfer.

Neuromorphic hardware tends to pose limits on the connectivity of deep networks that one can run on them. But also generic hardware and software implementations of deep learning run more efficiently for sparse networks. Several methods exist for pruning connections of a neural network after it was trained without connectivity constraints. We present an algorithm, DEEP R, that enables us to train directly a sparsely connected neural network. DEEP R automatically rewires the network during supervised training so that connections are there where they are most needed for the task, while its total number is all the time strictly bounded. We demonstrate that DEEP R can be used to train very sparse feedforward and recurrent neural networks on standard benchmark tasks with just a minor loss in performance. DEEP R is based on a rigorous theoretical foundation that views rewiring as stochastic sampling of network configurations from a posterior.

The paper presents Deep Rewiring, an algorithm that can be used to train deep neural networks when the network connectivity is severely constrained during training.

Deep learning's success has led to larger and larger models to handle more and more complex tasks; trained models can contain millions of parameters. These large models are compute- and memory-intensive, which makes it a challenge to deploy them with minimized latency, throughput, and storage requirements. Some model compression methods have been successfully applied on image classification and detection or language models, but there has been very little work compressing generative adversarial networks (GANs) performing complex tasks. In this paper, we show that a standard model compression technique, weight pruning, cannot be applied to GANs using existing methods. We then develop a self-supervised compression technique which uses the trained discriminator to supervise the training of a compressed generator. We show that this framework has a compelling performance to high degrees of sparsity, generalizes well to new tasks and models, and enables meaningful comparisons between different pruning granularities.

Existing pruning methods fail when applied to GANs tackling complex tasks, so we present a simple and robust method to prune generators that works well for a wide variety of networks and tasks.

Large-scale distributed training requires significant communication bandwidth for gradient exchange that limits the scalability of multi-node training, and requires expensive high-bandwidth network infrastructure. The situation gets even worse with distributed training on mobile devices (federated learning), which suffers from higher latency, lower throughput, and intermittent poor connections. In this paper, we find 99.9% of the gradient exchange in distributed SGD is redundant, and propose Deep Gradient Compression (DGC) to greatly reduce the communication bandwidth. To preserve accuracy during compression, DGC employs four methods: momentum correction, local gradient clipping, momentum factor masking, and warm-up training. We have applied Deep Gradient Compression to image classification, speech recognition, and language modeling with multiple datasets including Cifar10, ImageNet, Penn Treebank, and Librispeech Corpus. On these scenarios, Deep Gradient Compression achieves a gradient compression ratio from 270x to 600x without losing accuracy, cutting the gradient size of ResNet-50 from 97MB to 0.35MB, and for DeepSpeech from 488MB to 0.74MB. Deep gradient compression enables large-scale distributed training on inexpensive commodity 1Gbps Ethernet and facilitates distributed training on mobile.

we find 99.9% of the gradient exchange in distributed SGD is redundant; we reduce the communication bandwidth by two orders of magnitude without losing accuracy.

Image-to-image translation has recently received significant attention due to advances in deep learning. Most works focus on learning either a one-to-one mapping in an unsupervised way or a many-to-many mapping in a supervised way. However, a more practical setting is many-to-many mapping in an unsupervised way, which is harder due to the lack of supervision and the complex inner- and cross-domain variations. To alleviate these issues, we propose the Exemplar Guided & Semantically Consistent Image-to-image Translation (EGSC-IT) network which conditions the translation process on an exemplar image in the target domain. We assume that an image comprises of a content component which is shared across domains, and a style component specific to each domain. Under the guidance of an exemplar from the target domain we apply Adaptive Instance Normalization to the shared content component, which allows us to transfer the style information of the target domain to the source domain. To avoid semantic inconsistencies during translation that naturally appear due to the large inner- and cross-domain variations, we introduce the concept of feature masks that provide coarse semantic guidance without requiring the use of any semantic labels. Experimental results on various datasets show that EGSC-IT does not only translate the source image to diverse instances in the target domain, but also preserves the semantic consistency during the process.

We propose the Exemplar Guided & Semantically Consistent Image-to-image Translation (EGSC-IT) network which conditions the translation process on an exemplar image in the target domain.

Deep neural networks can learn meaningful representations of data. However, these representations are hard to interpret. For example, visualizing a latent layer is generally only possible for at most three dimensions. Neural networks are able to learn and benefit from much higher dimensional representations but these are not visually interpretable because nodes have arbitrary ordering within a layer. Here, we utilize the ability of the human observer to identify patterns in structured representations to visualize higher dimensions. To do so, we propose a class of regularizations we call \textit{Graph Spectral Regularizations} that impose graph-structure on latent layers. This is achieved by treating activations as signals on a predefined graph and constraining those activations using graph filters, such as low pass and wavelet-like filters. This framework allows for any kind of graph as well as filter to achieve a wide range of structured regularizations depending on the inference needs of the data. First, we show a synthetic example that the graph-structured layer can reveal topological features of the data. Next, we show that a smoothing regularization can impose semantically consistent ordering of nodes when applied to capsule nets. Further, we show that the graph-structured layer, using wavelet-like spatially localized filters, can form localized receptive fields for improved image and biomedical data interpretation. In other words, the mapping between latent layer, neurons and the output space becomes clear due to the localization of the activations. Finally, we show that when structured as a grid, the representations create coherent images that allow for image-processing techniques such as convolutions.

Imposing graph structure on neural network layers for improved visual interpretability.

Text generation is ubiquitous in many NLP tasks, from summarization, to dialogue and machine translation. The dominant parametric approach is based on locally normalized models which predict one word at a time. While these work remarkably well, they are plagued by exposure bias due to the greedy nature of the generation process. In this work, we investigate un-normalized energy-based models (EBMs) which operate not at the token but at the sequence level. In order to make training tractable, we first work in the residual of a pretrained locally normalized language model and second we train using noise contrastive estimation. Furthermore, since the EBM works at the sequence level, we can leverage pretrained bi-directional contextual representations, such as BERT and RoBERTa. Our experiments on two large language modeling datasets show that residual EBMs yield lower perplexity compared to locally normalized baselines. Moreover, generation via importance sampling is very efficient and of higher quality than the baseline models according to human evaluation.

We show that Energy-Based models when trained on the residual of an auto-regressive language model can be used effectively and efficiently to generate text.

We investigate the robustness properties of image recognition models equipped with two features inspired by human vision, an explicit episodic memory and a shape bias, at the ImageNet scale. As reported in previous work, we show that an explicit episodic memory improves the robustness of image recognition models against small-norm adversarial perturbations under some threat models. It does not, however, improve the robustness against more natural, and typically larger, perturbations. Learning more robust features during training appears to be necessary for robustness in this second sense. We show that features derived from a model that was encouraged to learn global, shape-based representations (Geirhos et al., 2019) do not only improve the robustness against natural perturbations, but when used in conjunction with an episodic memory, they also provide additional robustness against adversarial perturbations. Finally, we address three important design choices for the episodic memory: memory size, dimensionality of the memories and the retrieval method. We show that to make the episodic memory more compact, it is preferable to reduce the number of memories by clustering them, instead of reducing their dimensionality.

systematic study of large-scale cache-based image recognition models, focusing particularly on their robustness properties

Group convolutional neural networks (G-CNNs) can be used to improve classical CNNs by equipping them with the geometric structure of groups. Central in the success of G-CNNs is the lifting of feature maps to higher dimensional disentangled representations, in which data characteristics are effectively learned, geometric data-augmentations are made obsolete, and predictable behavior under geometric transformations (equivariance) is guaranteed via group theory. Currently, however, the practical implementations of G-CNNs are limited to either discrete groups (that leave the grid intact) or continuous compact groups such as rotations (that enable the use of Fourier theory). In this paper we lift these limitations and propose a modular framework for the design and implementation of G-CNNs for arbitrary Lie groups. In our approach the differential structure of Lie groups is used to expand convolution kernels in a generic basis of B-splines that is defined on the Lie algebra. This leads to a flexible framework that enables localized, atrous, and deformable convolutions in G-CNNs by means of respectively localized, sparse and non-uniform B-spline expansions. The impact and potential of our approach is studied on two benchmark datasets: cancer detection in histopathology slides (PCam dataset) in which rotation equivariance plays a key role and facial landmark localization (CelebA dataset) in which scale equivariance is important. In both cases, G-CNN architectures outperform their classical 2D counterparts and the added value of atrous and localized group convolutions is studied in detail.

The paper describes a flexible framework for building CNNs that are equivariant to a large class of transformations groups.

Global feature pooling is a modern variant of feature pooling providing better interpretatability and regularization. Although alternative pooling methods exist (eg. max, lp norm, stochastic), the averaging operation is still the dominating global pooling scheme in popular models. As fine-grained recognition requires learning subtle, discriminative features, we consider the question: is average pooling the optimal strategy? We first ask: ``is there a difference between features learned by global average and max pooling?'' Visualization and quantitative analysis show that max pooling encourages learning features of different spatial scales. We then ask ``is there a single global feature pooling variant that's most suitable for fine-grained recognition?'' A thorough evaluation of nine representative pooling algorithms finds that: max pooling outperforms average pooling consistently across models, datasets, and image resolutions; it does so by reducing the generalization gap; and generalized pooling's performance increases almost monotonically as it changes from average to max. We finally ask: ``what's the best way to combine two heterogeneous pooling schemes?'' Common strategies struggle because of potential gradient conflict but the ``freeze-and-train'' trick works best. We also find that post-global batch normalization helps with faster convergence and improves model performance consistently.

A benchmark of nine representative global pooling schemes reveals some interesting findings.

We present a technique to improve the generalization of deep representations learned on small labeled datasets by introducing self-supervised tasks as auxiliary loss functions. Although recent research has shown benefits of self-supervised learning (SSL) on large unlabeled datasets, its utility on small datasets is unknown. We find that SSL reduces the relative error rate of few-shot meta-learners by 4%-27%, even when the datasets are small and only utilizing images within the datasets. The improvements are greater when the training set is smaller or the task is more challenging. Though the benefits of SSL may increase with larger training sets, we observe that SSL can have a negative impact on performance when there is a domain shift between distribution of images used for meta-learning and SSL. Based on this analysis we present a technique that automatically select images for SSL from a large, generic pool of unlabeled images for a given dataset using a domain classifier that provides further improvements. We present results using several meta-learners and self-supervised tasks across datasets with varying degrees of domain shifts and label sizes to characterize the effectiveness of SSL for few-shot learning.

Self-supervision improves few-shot recognition on small and challenging datasets without relying on extra data; Extra data helps only when it is from the same or similar domain.

Abstraction of Markov Decision Processes is a useful tool for solving complex problems, as it can ignore unimportant aspects of an environment, simplifying the process of learning an optimal policy. In this paper, we propose a new algorithm for finding abstract MDPs in environments with continuous state spaces. It is based on MDP homomorphisms, a structure-preserving mapping between MDPs. We demonstrate our algorithm's ability to learns abstractions from collected experience and show how to reuse the abstractions to guide exploration in new tasks the agent encounters. Our novel task transfer method beats a baseline based on a deep Q-network.

We create abstract models of environments from experience and use them to learn new tasks faster.

A number of recent methods to understand neural networks have focused on quantifying the role of individual features. One such method, NetDissect identifies interpretable features of a model using the Broden dataset of visual semantic labels (colors, materials, textures, objects and scenes). Given the recent rise of a number of action recognition datasets, we propose extending the Broden dataset to include actions to better analyze learned action models. We describe the annotation process, results from interpreting action recognition models on the extended Broden dataset and examine interpretable feature paths to help us understand the conceptual hierarchy used to classify an action.

We expand Network Dissection to include action interpretation and examine interpretable feature paths to understand the conceptual hierarchy used to classify an action.

Automatic melody generation for pop music has been a long-time aspiration for both AI researchers and musicians. However, learning to generate euphonious melody has turned out to be highly challenging due to a number of factors. Representation of multivariate property of notes has been one of the primary challenges. It is also difficult to remain in the permissible spectrum of musical variety, outside of which would be perceived as a plain random play without auditory pleasantness. Observing the conventional structure of pop music poses further challenges. In this paper, we propose to represent each note and its properties as a unique ‘word,’ thus lessening the prospect of misalignments between the properties, as well as reducing the complexity of learning. We also enforce regularization policies on the range of notes, thus encouraging the generated melody to stay close to what humans would find easy to follow. Furthermore, we generate melody conditioned on song part information, thus replicating the overall structure of a full song. Experimental results demonstrate that our model can generate auditorily pleasant songs that are more indistinguishable from human-written ones than previous models.

We propose a novel model to represent notes and their properties, which can enhance the automatic melody generation.

Depth is a key component of Deep Neural Networks (DNNs), however, designing depth is heuristic and requires many human efforts. We propose AutoGrow to automate depth discovery in DNNs: starting from a shallow seed architecture, AutoGrow grows new layers if the growth improves the accuracy; otherwise, stops growing and thus discovers the depth. We propose robust growing and stopping policies to generalize to different network architectures and datasets. Our experiments show that by applying the same policy to different network architectures, AutoGrow can always discover near-optimal depth on various datasets of MNIST, FashionMNIST, SVHN, CIFAR10, CIFAR100 and ImageNet. For example, in terms of accuracy-computation trade-off, AutoGrow discovers a better depth combination in ResNets than human experts. Our AutoGrow is efficient. It discovers depth within similar time of training a single DNN.

A method that automatically grows layers in neural networks to discover optimal depth.

Given the importance of remote sensing, surprisingly little attention has been paid to it by the representation learning community. To address it and to speed up innovation in this domain, we provide simplified access to 5 diverse remote sensing datasets in a standardized form. We specifically explore in-domain representation learning and address the question of "what characteristics should a dataset have to be a good source for remote sensing representation learning". The established baselines achieve state-of-the-art performance on these datasets.

Exploration of in-domain representation learning for remote sensing datasets.

Generative seq2seq dialogue systems are trained to predict the next word in dialogues that have already occurred. They can learn from large unlabeled conversation datasets, build a deep understanding of conversational context, and generate a wide variety of responses. This flexibility comes at the cost of control. Undesirable responses in the training data will be reproduced by the model at inference time, and longer generations often don’t make sense. Instead of generating responses one word at a time, we train a classifier to choose from a predefined list of full responses. The classifier is trained on (conversation context, response class) pairs, where each response class is a noisily labeled group of interchangeable responses. At inference, we generate the exemplar response associated with the predicted response class. Experts can edit and improve these exemplar responses over time without retraining the classifier or invalidating old training data. Human evaluation of 775 unseen doctor/patient conversations shows that this tradeoff improves responses. Only 12% of our discriminative approach’s responses are worse than the doctor’s response in the same conversational context, compared to 18% for the generative model. A discriminative model trained without any manual labeling of response classes achieves equal performance to the generative model.

Avoid generating responses one word at a time by using weak supervision to training a classifier to pick a full response.

There is a previously identified equivalence between wide fully connected neural networks (FCNs) and Gaussian processes (GPs). This equivalence enables, for instance, test set predictions that would have resulted from a fully Bayesian, infinitely wide trained FCN to be computed without ever instantiating the FCN, but by instead evaluating the corresponding GP. In this work, we derive an analogous equivalence for multi-layer convolutional neural networks (CNNs) both with and without pooling layers, and achieve state of the art results on CIFAR10 for GPs without trainable kernels. We also introduce a Monte Carlo method to estimate the GP corresponding to a given neural network architecture, even in cases where the analytic form has too many terms to be computationally feasible. Surprisingly, in the absence of pooling layers, the GPs corresponding to CNNs with and without weight sharing are identical. As a consequence, translation equivariance, beneficial in finite channel CNNs trained with stochastic gradient descent (SGD), is guaranteed to play no role in the Bayesian treatment of the infinite channel limit - a qualitative difference between the two regimes that is not present in the FCN case. We confirm experimentally, that while in some scenarios the performance of SGD-trained finite CNNs approaches that of the corresponding GPs as the channel count increases, with careful tuning SGD-trained CNNs can significantly outperform their corresponding GPs, suggesting advantages from SGD training compared to fully Bayesian parameter estimation.

Finite-width SGD trained CNNs vs. infinitely wide fully Bayesian CNNs. Who wins?

Bayesian inference promises to ground and improve the performance of deep neural networks. It promises to be robust to overfitting, to simplify the training procedure and the space of hyperparameters, and to provide a calibrated measure of uncertainty that can enhance decision making, agent exploration and prediction fairness. Markov Chain Monte Carlo (MCMC) methods enable Bayesian inference by generating samples from the posterior distribution over model parameters. Despite the theoretical advantages of Bayesian inference and the similarity between MCMC and optimization methods, the performance of sampling methods has so far lagged behind optimization methods for large scale deep learning tasks. We aim to fill this gap and introduce ATMC, an adaptive noise MCMC algorithm that estimates and is able to sample from the posterior of a neural network. ATMC dynamically adjusts the amount of momentum and noise applied to each parameter update in order to compensate for the use of stochastic gradients. We use a ResNet architecture without batch normalization to test ATMC on the Cifar10 benchmark and the large scale ImageNet benchmark and show that, despite the absence of batch normalization, ATMC outperforms a strong optimization baseline in terms of both classification accuracy and test log-likelihood. We show that ATMC is intrinsically robust to overfitting on the training data and that ATMC provides a better calibrated measure of uncertainty compared to the optimization baseline.

We scale Bayesian Inference to ImageNet classification and achieve competitive results accuracy and uncertainty calibration.

Now GANs can generate more and more realistic face images that can easily fool human beings. In contrast, a common convolutional neural network(CNN), e.g. ResNet-18, can achieve more than 99.9% accuracy in discerning fake/real faces if training and testing faces are from the same source. In this paper, we performed both human studies and CNN experiments, which led us to two important findings. One finding is that the textures of fake faces are substantially different from real ones. CNNs can capture local image texture information for recognizing fake/real face, while such cues are easily overlooked by humans. The other finding is that global image texture information is more robust to image editing and generalizable to fake faces from different GANs and datasets. Based on the above findings, we propose a novel architecture coined as Gram-Net, which incorporates “Gram Block” in multiple semantic levels to extract global image texture representations. Experimental results demonstrate that our Gram-Net performs better than existing approaches for fake face detection. Especially, our Gram-Net is more robust to image editing, e.g. downsampling, JPEG compression, blur, and noise. More importantly, our Gram-Net generalizes significantly better in detecting fake faces from GAN models not seen in the training phase.

An empirical study on fake images reveals that texture is an important cue that current fake images differ from real images. Our improved model capturing global texture statistics shows better cross-GAN fake image detection performance.

The Wasserstein probability metric has received much attention from the machine learning community. Unlike the Kullback-Leibler divergence, which strictly measures change in probability, the Wasserstein metric reflects the underlying geometry between outcomes. The value of being sensitive to this geometry has been demonstrated, among others, in ordinal regression and generative modelling, and most recently in reinforcement learning. In this paper we describe three natural properties of probability divergences that we believe reflect requirements from machine learning: sum invariance, scale sensitivity, and unbiased sample gradients. The Wasserstein metric possesses the first two properties but, unlike the Kullback-Leibler divergence, does not possess the third. We provide empirical evidence suggesting this is a serious issue in practice. Leveraging insights from probabilistic forecasting we propose an alternative to the Wasserstein metric, the Cramér distance. We show that the Cramér distance possesses all three desired properties, combining the best of the Wasserstein and Kullback-Leibler divergences. We give empirical results on a number of domains comparing these three divergences. To illustrate the practical relevance of the Cramér distance we design a new algorithm, the Cramér Generative Adversarial Network (GAN), and show that it has a number of desirable properties over the related Wasserstein GAN.

The Wasserstein distance is hard to minimize with stochastic gradient descent, while the Cramer distance can be optimized easily and works just as well.

We humans have an innate understanding of the asymmetric progression of time, which we use to efficiently and safely perceive and manipulate our environment. Drawing inspiration from that, we approach the problem of learning an arrow of time in a Markov (Decision) Process. We illustrate how a learned arrow of time can capture salient information about the environment, which in turn can be used to measure reachability, detect side-effects and to obtain an intrinsic reward signal. Finally, we propose a simple yet effective algorithm to parameterize the problem at hand and learn an arrow of time with a function approximator (here, a deep neural network). Our empirical results span a selection of discrete and continuous environments, and demonstrate for a class of stochastic processes that the learned arrow of time agrees reasonably well with a well known notion of an arrow of time due to Jordan, Kinderlehrer and Otto (1998).

We learn the arrow of time for MDPs and use it to measure reachability, detect side-effects and obtain a curiosity reward signal.

We formulate stochastic gradient descent (SGD) as a novel factorised Bayesian filtering problem, in which each parameter is inferred separately, conditioned on the corresopnding backpropagated gradient. Inference in this setting naturally gives rise to BRMSprop and BAdam: Bayesian variants of RMSprop and Adam. Remarkably, the Bayesian approach recovers many features of state-of-the-art adaptive SGD methods, including amongst others root-mean-square normalization, Nesterov acceleration and AdamW. As such, the Bayesian approach provides one explanation for the empirical effectiveness of state-of-the-art adaptive SGD algorithms. Empirically comparing BRMSprop and BAdam with naive RMSprop and Adam on MNIST, we find that Bayesian methods have the potential to considerably reduce test loss and classification error.

We formulated SGD as a Bayesian filtering problem, and show that this gives rise to RMSprop, Adam, AdamW, NAG and other features of state-of-the-art adaptive methods

Data augmentation (DA) has been widely utilized to improve generalization in training deep neural networks. Recently, human-designed data augmentation has been gradually replaced by automatically learned augmentation policy. Through finding the best policy in well-designed search space of data augmentation, AutoAugment (Cubuk et al., 2019) can significantly improve validation accuracy on image classification tasks. However, this approach is not computationally practical for large-scale problems. In this paper, we develop an adversarial method to arrive at a computationally-affordable solution called Adversarial AutoAugment, which can simultaneously optimize target related object and augmentation policy search loss. The augmentation policy network attempts to increase the training loss of a target network through generating adversarial augmentation policies, while the target network can learn more robust features from harder examples to improve the generalization. In contrast to prior work, we reuse the computation in target network training for policy evaluation, and dispense with the retraining of the target network. Compared to AutoAugment, this leads to about 12x reduction in computing cost and 11x shortening in time overhead on ImageNet. We show experimental results of our approach on CIFAR-10/CIFAR-100, ImageNet, and demonstrate significant performance improvements over state-of-the-art. On CIFAR-10, we achieve a top-1 test error of 1.36%, which is the currently best performing single model. On ImageNet, we achieve a leading performance of top-1 accuracy 79.40% on ResNet-50 and 80.00% on ResNet-50-D without extra data.

We introduce the idea of adversarial learning into automatic data augmentation to improve the generalization of a targe network.

In this study we focus on first-order meta-learning algorithms that aim to learn a parameter initialization of a network which can quickly adapt to new concepts, given a few examples. We investigate two approaches to enhance generalization and speed of learning of such algorithms, particularly expanding on the Reptile (Nichol et al., 2018) algorithm. We introduce a novel regularization technique called meta-step gradient pruning and also investigate the effects of increasing the depth of network architectures in first-order meta-learning. We present an empirical evaluation of both approaches, where we match benchmark few-shot image classification results with 10 times fewer iterations using Mini-ImageNet dataset and with the use of deeper networks, we attain accuracies that surpass the current benchmarks of few-shot image classification using Omniglot dataset.

The study introduces two approaches to enhance generalization of first-order meta-learning and presents empirical evaluation on few-shot image classification.

In this paper, we propose the use of in-training matrix factorization to reduce the model size for neural machine translation. Using in-training matrix factorization, parameter matrices may be decomposed into the products of smaller matrices, which can compress large machine translation architectures by vastly reducing the number of learnable parameters. We apply in-training matrix factorization to different layers of standard neural architectures and show that in-training factorization is capable of reducing nearly 50% of learnable parameters without any associated loss in BLEU score. Further, we find that in-training matrix factorization is especially powerful on embedding layers, providing a simple and effective method to curtail the number of parameters with minimal impact on model performance, and, at times, an increase in performance.

This paper proposes using matrix factorization at training time for neural machine translation, which can reduce model size and decrease training time without harming performance.

Though state-of-the-art sentence representation models can perform tasks requiring significant knowledge of grammar, it is an open question how best to evaluate their grammatical knowledge. We explore five experimental methods inspired by prior work evaluating pretrained sentence representation models. We use a single linguistic phenomenon, negative polarity item (NPI) licensing, as a case study for our experiments. NPIs like 'any' are grammatical only if they appear in a licensing environment like negation ('Sue doesn't have any cats' vs. '*Sue has any cats'). This phenomenon is challenging because of the variety of NPI licensing environments that exist. We introduce an artificially generated dataset that manipulates key features of NPI licensing for the experiments. We find that BERT has significant knowledge of these features, but its success varies widely across different experimental methods. We conclude that a variety of methods is necessary to reveal all relevant aspects of a model's grammatical knowledge in a given domain.

Different methods for analyzing BERT suggest different (but compatible) conclusions in a case study on NPIs.

The primate visual system builds robust, multi-purpose representations of the external world in order to support several diverse downstream cortical processes. Such representations are required to be invariant to the sensory inconsistencies caused by dynamically varying lighting, local texture distortion, etc. A key architectural feature combating such environmental irregularities is ‘long-range horizontal connections’ that aid the perception of the global form of objects. In this work, we explore the introduction of such horizontal connections into standard deep convolutional networks; we present V1Net -- a novel convolutional-recurrent unit that models linear and nonlinear horizontal inhibitory and excitatory connections inspired by primate visual cortical connectivity. We introduce the Texturized Challenge -- a new benchmark to evaluate object recognition performance under perceptual noise -- which we use to evaluate V1Net against an array of carefully selected control models with/without recurrent processing. Additionally, we present results from an ablation study of V1Net demonstrating the utility of diverse neurally inspired horizontal connections for state-of-the-art AI systems on the task of object boundary detection from natural images. We also present the emergence of several biologically plausible horizontal connectivity patterns, namely center-on surround-off, association fields and border-ownership connectivity patterns in a V1Net model trained to perform boundary detection on natural images from the Berkeley Segmentation Dataset 500 (BSDS500). Our findings suggest an increased representational similarity between V1Net and biological visual systems, and highlight the importance of neurally inspired recurrent contextual processing principles for learning visual representations that are robust to perceptual noise and furthering the state-of-the-art in computer vision.

In this work, we present V1Net -- a novel recurrent neural network modeling cortical horizontal connections that give rise to robust visual representations through perceptual grouping.

Humans understand novel sentences by composing meanings and roles of core language components. In contrast, neural network models for natural language modeling fail when such compositional generalization is required. The main contribution of this paper is to hypothesize that language compositionality is a form of group-equivariance. Based on this hypothesis, we propose a set of tools for constructing equivariant sequence-to-sequence models. Throughout a variety of experiments on the SCAN tasks, we analyze the behavior of existing models under the lens of equivariance, and demonstrate that our equivariant architecture is able to achieve the type compositional generalization required in human language understanding.

We propose a link between permutation equivariance and compositional generalization, and provide equivariant language models

Variational inference (VI) is a popular approach for approximate Bayesian inference that is particularly promising for highly parameterized models such as deep neural networks. A key challenge of variational inference is to approximate the posterior over model parameters with a distribution that is simpler and tractable yet sufficiently expressive. In this work, we propose a method for training highly flexible variational distributions by starting with a coarse approximation and iteratively refining it. Each refinement step makes cheap, local adjustments and only requires optimization of simple variational families. We demonstrate theoretically that our method always improves a bound on the approximation (the Evidence Lower BOund) and observe this empirically across a variety of benchmark tasks. In experiments, our method consistently outperforms recent variational inference methods for deep learning in terms of log-likelihood and the ELBO. We see that the gains are further amplified on larger scale models, significantly outperforming standard VI and deep ensembles on residual networks on CIFAR10.

The paper proposes an algorithm to increase the flexibility of the variational posterior in Bayesian neural networks through iterative optimization.

In this paper, we propose a residual non-local attention network for high-quality image restoration. Without considering the uneven distribution of information in the corrupted images, previous methods are restricted by local convolutional operation and equal treatment of spatial- and channel-wise features. To address this issue, we design local and non-local attention blocks to extract features that capture the long-range dependencies between pixels and pay more attention to the challenging parts. Specifically, we design trunk branch and (non-)local mask branch in each (non-)local attention block. The trunk branch is used to extract hierarchical features. Local and non-local mask branches aim to adaptively rescale these hierarchical features with mixed attentions. The local mask branch concentrates on more local structures with convolutional operations, while non-local attention considers more about long-range dependencies in the whole feature map. Furthermore, we propose residual local and non-local attention learning to train the very deep network, which further enhance the representation ability of the network. Our proposed method can be generalized for various image restoration applications, such as image denoising, demosaicing, compression artifacts reduction, and super-resolution. Experiments demonstrate that our method obtains comparable or better results compared with recently leading methods quantitatively and visually.

New state-of-the-art framework for image restoration

Most approaches to learning action planning models heavily rely on a significantly large volume of training samples or plan observations. In this paper, we adopt a different approach based on deductive learning from domain-specific knowledge, specifically from logic formulae that specify constraints about the possible states of a given domain. The minimal input observability required by our approach is a single example composed of a full initial state and a partial goal state. We will show that exploiting specific domain knowledge enables to constrain the space of possible action models as well as to complete partial observations, both of which turn out helpful to learn good-quality action models.

Hybrid approach to model acquisition that compensates a lack of available data with domain specific knowledge provided by experts

We release the largest public ECG dataset of continuous raw signals for representation learning containing over 11k patients and 2 billion labelled beats. Our goal is to enable semi-supervised ECG models to be made as well as to discover unknown subtypes of arrhythmia and anomalous ECG signal events. To this end, we propose an unsupervised representation learning task, evaluated in a semi-supervised fashion. We provide a set of baselines for different feature extractors that can be built upon. Additionally, we perform qualitative evaluations on results from PCA embeddings, where we identify some clustering of known subtypes indicating the potential for representation learning in arrhythmia sub-type discovery.

We release a dataset constructed from single-lead ECG data from 11,000 patients who were prescribed to use the {DEVICENAME}(TM) device.

As the basic building block of Convolutional Neural Networks (CNNs), the convolutional layer is designed to extract local patterns and lacks the ability to model global context in its nature. Many efforts have been recently made to complement CNNs with the global modeling ability, especially by a family of works on global feature interaction. In these works, the global context information is incorporated into local features before they are fed into convolutional layers. However, research on neuroscience reveals that, besides influences changing the inputs to our neurons, the neurons' ability of modifying their functions dynamically according to context is essential for perceptual tasks, which has been overlooked in most of CNNs. Motivated by this, we propose one novel Context-Gated Convolution (CGC) to explicitly modify the weights of convolutional layers adaptively under the guidance of global context. As such, being aware of the global context, the modulated convolution kernel of our proposed CGC can better extract representative local patterns and compose discriminative features. Moreover, our proposed CGC is lightweight, amenable to modern CNN architectures, and consistently improves the performance of CNNs according to extensive experiments on image classification, action recognition, and machine translation.

A novel Context-Gated Convolution which incorporates global context information into CNNs by explicitly modulating convolution kernels, and thus captures more representative local patterns and extract discriminative features.

We analyze the trade-off between quantization noise and clipping distortion in low precision networks. We identify the statistics of various tensors, and derive exact expressions for the mean-square-error degradation due to clipping. By optimizing these expressions, we show marked improvements over standard quantization schemes that normally avoid clipping. For example, just by choosing the accurate clipping values, more than 40\% accuracy improvement is obtained for the quantization of VGG-16 to 4-bits of precision. Our results have many applications for the quantization of neural networks at both training and inference time.

We analyze the trade-off between quantization noise and clipping distortion in low precision networks, and show marked improvements over standard quantization schemes that normally avoid clipping

Batch Normalization (BN) is one of the most widely used techniques in Deep Learning field. But its performance can awfully degrade with insufficient batch size. This weakness limits the usage of BN on many computer vision tasks like detection or segmentation, where batch size is usually small due to the constraint of memory consumption. Therefore many modified normalization techniques have been proposed, which either fail to restore the performance of BN completely, or have to introduce additional nonlinear operations in inference procedure and increase huge consumption. In this paper, we reveal that there are two extra batch statistics involved in backward propagation of BN, on which has never been well discussed before. The extra batch statistics associated with gradients also can severely affect the training of deep neural network. Based on our analysis, we propose a novel normalization method, named Moving Average Batch Normalization (MABN). MABN can completely restore the performance of vanilla BN in small batch cases, without introducing any additional nonlinear operations in inference procedure. We prove the benefits of MABN by both theoretical analysis and experiments. Our experiments demonstrate the effectiveness of MABN in multiple computer vision tasks including ImageNet and COCO. The code has been released in https://github.com/megvii-model/MABN.

We propose a novel normalization method to handle small batch size cases.

We present a simple proof for the benefit of depth in multi-layer feedforward network with rectifed activation (``"depth separation"). Specifically we present a sequence of classification problems f_i such that (a) for any fixed depth rectified network we can find an index m such that problems with index > m require exponential network width to fully represent the function f_m; and (b) for any problem f_m in the family, we present a concrete neural network with linear depth and bounded width that fully represents it. While there are several previous work showing similar results, our proof uses substantially simpler tools and techniques, and should be accessible to undergraduate students in computer science and people with similar backgrounds.

ReLU MLP depth seperation proof with gemoteric arguments

The rich and accessible labeled data fuel the revolutionary success of deep learning. Nonetheless, massive supervision remains a luxury for many real applications, boosting great interest in label-scarce techniques such as few-shot learning (FSL). An intuitively feasible approach to FSL is to conduct data augmentation via synthesizing additional training samples. The key to this approach is how to guarantee both discriminability and diversity of the synthesized samples. In this paper, we propose a novel FSL model, called $\textrm{D}^2$GAN, which synthesizes Diverse and Discriminative features based on Generative Adversarial Networks (GAN). $\textrm{D}^2$GAN secures discriminability of the synthesized features by constraining them to have high correlation with real features of the same classes while low correlation with those of different classes. Based on the observation that noise vectors that are closer in the latent code space are more likely to be collapsed into the same mode when mapped to feature space, $\textrm{D}^2$GAN incorporates a novel anti-collapse regularization term, which encourages feature diversity by penalizing the ratio of the logarithmic similarity of two synthesized features and the logarithmic similarity of the latent codes generating them. Experiments on three common benchmark datasets verify the effectiveness of $\textrm{D}^2$GAN by comparing with the state-of-the-art.

A new GAN based few-shot learning algorithm by synthesizing diverse and discriminative Features

The lack of crisp mathematical models that capture the structure of real-world data sets is a major obstacle to the detailed theoretical understanding of deep neural networks. Here, we first demonstrate the effect of structured data sets by experimentally comparing the dynamics and the performance of two-layer networks trained on two different data sets: (i) an unstructured synthetic data set containing random i.i.d. inputs, and (ii) a simple canonical data set such as MNIST images. Our analysis reveals two phenomena related to the dynamics of the networks and their ability to generalise that only appear when training on structured data sets. Second, we introduce a generative model for data sets, where high-dimensional inputs lie on a lower-dimensional manifold and have labels that depend only on their position within this manifold. We call it the *hidden manifold model* and we experimentally demonstrate that training networks on data sets drawn from this model reproduces both the phenomena seen during training on MNIST.

We demonstrate how structure in data sets impacts neural networks and introduce a generative model for synthetic data sets that reproduces this impact.

In this paper, we study deep diagonal circulant neural networks, that is deep neural networks in which weight matrices are the product of diagonal and circulant ones. Besides making a theoretical analysis of their expressivity, we introduced principled techniques for training these models: we devise an initialization scheme and proposed a smart use of non-linearity functions in order to train deep diagonal circulant networks. Furthermore, we show that these networks outperform recently introduced deep networks with other types of structured layers. We conduct a thorough experimental study to compare the performance of deep diagonal circulant networks with state of the art models based on structured matrices and with dense models. We show that our models achieve better accuracy than other structured approaches while required 2x fewer weights as the next best approach. Finally we train deep diagonal circulant networks to build a compact and accurate models on a real world video classification dataset with over 3.8 million training examples.

We train deep neural networks based on diagonal and circulant matrices, and show that this type of networks are both compact and accurate on real world applications.

Interpretability has largely focused on local explanations, i.e. explaining why a model made a particular prediction for a sample. These explanations are appealing due to their simplicity and local fidelity. However, they do not provide information about the general behavior of the model. We propose to leverage model distillation to learn global additive explanations that describe the relationship between input features and model predictions. These global explanations take the form of feature shapes, which are more expressive than feature attributions. Through careful experimentation, we show qualitatively and quantitatively that global additive explanations are able to describe model behavior and yield insights about models such as neural nets. A visualization of our approach applied to a neural net as it is trained is available at https://youtu.be/ErQYwNqzEdc

We propose to leverage model distillation to learn global additive explanations in the form of feature shapes (that are more expressive than feature attributions) for models such as neural nets trained on tabular data.

A lot of the recent success in natural language processing (NLP) has been driven by distributed vector representations of words trained on large amounts of text in an unsupervised manner. These representations are typically used as general purpose features for words across a range of NLP problems. However, extending this success to learning representations of sequences of words, such as sentences, remains an open problem. Recent work has explored unsupervised as well as supervised learning techniques with different training objectives to learn general purpose fixed-length sentence representations. In this work, we present a simple, effective multi-task learning framework for sentence representations that combines the inductive biases of diverse training objectives in a single model. We train this model on several data sources with multiple training objectives on over 100 million sentences. Extensive experiments demonstrate that sharing a single recurrent sentence encoder across weakly related tasks leads to consistent improvements over previous methods. We present substantial improvements in the context of transfer learning and low-resource settings using our learned general-purpose representations.

A large-scale multi-task learning framework with diverse training objectives to learn fixed-length sentence representations

In a time where neural networks are increasingly adopted in sensitive applications, algorithmic bias has emerged as an issue with moral implications. While there are myriad ways that a system may be compromised by bias, systematically isolating and evaluating existing systems on such scenarios is non-trivial, i.e., bias may be subtle, natural and inherently difficult to quantify. To this end, this paper proposes the first systematic study of benchmarking state-of-the-art neural models against biased scenarios. More concretely, we postulate that the bias annotator problem can be approximated with neural models, i.e., we propose generative models of latent bias to deliberately and unfairly associate latent features to a specific class. All in all, our framework provides a new way for principled quantification and evaluation of models against biased datasets. Consequently, we find that state-of-the-art NLP models (e.g., BERT, RoBERTa, XLNET) are readily compromised by biased data.

We propose a neural bias annotator to benchmark models on their robustness to biased text datasets.

We consider the problem of topic modeling in a weakly semi-supervised setting. In this scenario, we assume that the user knows a priori a subset of the topics she wants the model to learn and is able to provide a few exemplar documents for those topics. In addition, while each document may typically consist of multiple topics, we do not assume that the user will identify all its topics exhaustively. Recent state-of-the-art topic models such as NVDM, referred to herein as Neural Topic Models (NTMs), fall under the variational autoencoder framework. We extend NTMs to the weakly semi-supervised setting by using informative priors in the training objective. After analyzing the effect of informative priors, we propose a simple modification of the NVDM model using a logit-normal posterior that we show achieves better alignment to user-desired topics versus other NTM models.

We propose supervising VAE-style topic models by intelligently adjusting the prior on a per document basis. We find a logit-normal posterior provides the best performance.

Analyzing deep neural networks (DNNs) via information plane (IP) theory has gained tremendous attention recently as a tool to gain insight into, among others, their generalization ability. However, it is by no means obvious how to estimate mutual information (MI) between each hidden layer and the input/desired output, to construct the IP. For instance, hidden layers with many neurons require MI estimators with robustness towards the high dimensionality associated with such layers. MI estimators should also be able to naturally handle convolutional layers, while at the same time being computationally tractable to scale to large networks. None of the existing IP methods to date have been able to study truly deep Convolutional Neural Networks (CNNs), such as the e.g.\ VGG-16. In this paper, we propose an IP analysis using the new matrix--based R\'enyi's entropy coupled with tensor kernels over convolutional layers, leveraging the power of kernel methods to represent properties of the probability distribution independently of the dimensionality of the data. The obtained results shed new light on the previous literature concerning small-scale DNNs, however using a completely new approach. Importantly, the new framework enables us to provide the first comprehensive IP analysis of contemporary large-scale DNNs and CNNs, investigating the different training phases and providing new insights into the training dynamics of large-scale neural networks.

First comprehensive information plane analysis of large scale deep neural networks using matrix based entropy and tensor kernels.

Developing agents that can learn to follow natural language instructions has been an emerging research direction. While being accessible and flexible, natural language instructions can sometimes be ambiguous even to humans. To address this, we propose to utilize programs, structured in a formal language, as a precise and expressive way to specify tasks. We then devise a modular framework that learns to perform a task specified by a program – as different circumstances give rise to diverse ways to accomplish the task, our framework can perceive which circumstance it is currently under, and instruct a multitask policy accordingly to fulfill each subtask of the overall task. Experimental results on a 2D Minecraft environment not only demonstrate that the proposed framework learns to reliably accomplish program instructions and achieves zero-shot generalization to more complex instructions but also verify the efficiency of the proposed modulation mechanism for learning the multitask policy. We also conduct an analysis comparing various models which learn from programs and natural language instructions in an end-to-end fashion.

We propose a modular framework that can accomplish tasks specified by programs and achieve zero-shot generalization to more complex tasks.

We analyze the convergence of (stochastic) gradient descent algorithm for learning a convolutional filter with Rectified Linear Unit (ReLU) activation function. Our analysis does not rely on any specific form of the input distribution and our proofs only use the definition of ReLU, in contrast with previous works that are restricted to standard Gaussian input. We show that (stochastic) gradient descent with random initialization can learn the convolutional filter in polynomial time and the convergence rate depends on the smoothness of the input distribution and the closeness of patches. To the best of our knowledge, this is the first recovery guarantee of gradient-based algorithms for convolutional filter on non-Gaussian input distributions. Our theory also justifies the two-stage learning rate strategy in deep neural networks. While our focus is theoretical, we also present experiments that justify our theoretical findings.

We prove randomly initialized (stochastic) gradient descent learns a convolutional filter in polynomial time.

Deep neural networks (DNNs) are widely adopted in real-world cognitive applications because of their high accuracy. The robustness of DNN models, however, has been recently challenged by adversarial attacks where small disturbance on input samples may result in misclassification. State-of-the-art defending algorithms, such as adversarial training or robust optimization, improve DNNs' resilience to adversarial attacks by paying high computational costs. Moreover, these approaches are usually designed to defend one or a few known attacking techniques only. The effectiveness to defend other types of attacking methods, especially those that have not yet been discovered or explored, cannot be guaranteed. This work aims for a general approach of enhancing the robustness of DNN models under adversarial attacks. In particular, we propose Bamboo -- the first data augmentation method designed for improving the general robustness of DNN without any hypothesis on the attacking algorithms. Bamboo augments the training data set with a small amount of data uniformly sampled on a fixed radius ball around each training data and hence, effectively increase the distance between natural data points and decision boundary. Our experiments show that Bamboo substantially improve the general robustness against arbitrary types of attacks and noises, achieving better results comparing to previous adversarial training methods, robust optimization methods and other data augmentation methods with the same amount of data points.

The first data augmentation method specially designed for improving the general robustness of DNN without any hypothesis on the attacking algorithms.

The ability to synthesize realistic patterns of neural activity is crucial for studying neural information processing. Here we used the Generative Adversarial Networks (GANs) framework to simulate the concerted activity of a population of neurons. We adapted the Wasserstein-GAN variant to facilitate the generation of unconstrained neural population activity patterns while still benefiting from parameter sharing in the temporal domain. We demonstrate that our proposed GAN, which we termed Spike-GAN, generates spike trains that match accurately the first- and second-order statistics of datasets of tens of neurons and also approximates well their higher-order statistics. We applied Spike-GAN to a real dataset recorded from salamander retina and showed that it performs as well as state-of-the-art approaches based on the maximum entropy and the dichotomized Gaussian frameworks. Importantly, Spike-GAN does not require to specify a priori the statistics to be matched by the model, and so constitutes a more flexible method than these alternative approaches. Finally, we show how to exploit a trained Spike-GAN to construct 'importance maps' to detect the most relevant statistical structures present in a spike train. Spike-GAN provides a powerful, easy-to-use technique for generating realistic spiking neural activity and for describing the most relevant features of the large-scale neural population recordings studied in modern systems neuroscience.

Using Wasserstein-GANs to generate realistic neural activity and to detect the most relevant features present in neural population patterns.